BirdCLEF 2026 — acoustic species identification

Species identification from field audio recordings, combining an EfficientNet-B0 sound event detection model with Perch v2 embeddings. Final score: macro ROC-AUC of approximately 0.930.

The most transferable lesson had nothing to do with audio. Under a compute budget, only inference-side and infrastructure optimisations are safe: caching, batching, efficient loading. Cutting training epochs or reducing the number of folds looks like the same kind of saving and is not — it degrades the score reliably. I have applied that rule in every competition since.